Using sgrep for querying structured text files 1

advertisement
Using sgrep for querying structured text files1
Jani Jaakkola
Pekka Kilpeläinen
University of Helsinki
Department of Computer Science
P.O. Box 26
FIN-00014 UNIVERSITY OF HELSINKI
Finland
{jjaakkol,kilpelai}@cs.helsinki.fi
Abstract
Sgrep is a Unix tool for searching the contents of text files. Sgrep implements an algebra of unrestricted text fragments called regions. The algebra
allows the retrieval of document components, represented as regions, based
on conditions on their relative containment and ordering. This simple yet
powerful model is suitable for querying structured document formats like
electronic mail, RTF, LaTeX, HTML, or SGML documents. We describe
the sgrep query language and give examples of its use. Especially, we
explain how sgrep can be used for querying and assembling SGML documents.
Keywords: Text search tools, structured documents, SGML.
1 Introduction
Every-day data processing creates large amounts of structured documents that have a
more or less regular structure. Familiar examples of such structured documents include
electronic mail, Usenet news, program source code, HTML pages, and SGML documents. Users seldom bother to index or to store such volatile texts in advanced information retrieval systems, even that would support the eventual searching of the
documents based on their content and their structure. Therefore general purpose text
search tools, such as grep [ThR75], Awk [AKW88] and Perl [WaS91], are often
used to search structured text documents.
The drawback with ordinary text tools is that they do not support any high-level structures in the documents - they treat text only in units of rows, words and characters. Utilizing further hierarchical structure is either impossible or requires considerable
programming. There is obviously a need for simple tools that would make the structure
of text available for querying and manipulating. Sgrep (structured grep) [JK96]
1. Presented at SGML Finland 1996, Espoo, Finland, October 4-5, 1996.
Using sgrep for querying structured text files
5 November 1996
1
offers such a tool by implementing a query algebra for retrieving hierarchical text
structures.
The main power of sgrep is to treat text as unrestricted fragments called regions, which
are natural representations for components of, e.g., SGML documents. Sgrep retrieves
these regions from the text based on queries expressing relative containment and ordering conditions. The regions may nest and overlap arbitrarily, and neither their number
nor their length is restricted. Sgrep does not assume anything about the structure or
content of the files it searches. Because there are no “syntactically incorrect” documents for sgrep, it can be applied in a wide variety of tasks. Sgrep can be used to search
either C source programs, SGML documents or binary files. On the other hand, the
generality of sgrep also means that the user has to write queries carefully to match the
markup in the files.
The query language of sgrep is based on a region algebra which was first introduced by
Burkowski [Bur92] and later developed further by Clarke, Cormack and Burkowski
[CCB95]. The region algebra consists of operations similar to query languages in some
existing text retrieval systems such as PAT [SaT92]. Sgrep differs from its predecessors
in its ability to treat arbitrarily nesting text regions, which is necessary for manipulating nested document elements like lists within lists.
The rest of the paper is organized as follows. First we describe briefly the basic notions
of how sgrep models text in Section 2. Then we describe the query language of sgrep
using simple examples in Section 3. Section 4 describes how sgrep is used for manipulating SGML documents. Section 5 is a short conclusion.
2 Data model of sgrep
In this section we describe the data model of sgrep, i.e., the basic structures that sgrep
provides for manipulating text. The basic unit of data in sgrep is a region. A region is a
non-empty fragment of text identified by its start and end positions. A region can be
anything from a single character to the whole contents of a file. Sgrep operates on sets
of regions called region sets. A region set can contain any regions, but each region can
occur at most once. Often the regions in a set share some common property. For example, they could be document elements of the same type.
Region sets can be written in sgrep queries explicitly by enclosing in brackets [ ] a
sequence of constant regions denoted by pairs of indices (start, end). For an example,
assume that we have a file named file10 consisting of ten characters:
% cat file10
0123456789
Now the region (0,9) equals to the contents of the entire file. The regions (0,0) and (9,9)
comprise the first and the last character of the file, correspondingly. There are also
built-in expressions start and end for region sets consisting of the first and the last
character, correspondingly. Below is an example of evaluating a constant region set.
The -o option allows an output format for regions to be specified. In this case the contents of each region is output between two hyphens.
Using sgrep for querying structured text files
5 November 1996
2
% sgrep -o"-%r-" '[(0,5)(2,9)(3,8)(4,7)(4,9)]' file10
-012345--23456789--345678--4567--456789Regions expressed explicitly by indices, as above, are rarely used. Instead, regions
retrieved as occurrences of a text pattern are the normal starting point for an sgrep
query, as we see in the next section.
The manipulation of region sets described in the next section is based on conditions on
the relative containment and ordering of regions. Let a = (a.s, a.e) and b = (b.s, b.e) be
two regions. If b occurs properly within a, that is, if a.s < b.s <= b.e <= a.e or a.s <=
b.s <= b.e < a.e, we say that region a contains region b, that region b is included in
region a, and that the regions are nested. If region b starts after the end of region a,
region a precedes region b, region b follows region a, and the regions are disjoint. If
two regions are neither nested or disjoint, they are overlapping.
3 Sgrep query language
In this section we briefly describe the most central sgrep query language features.
Sgrep implements a fully compositional algebra of region sets, which means that the
result of any operation can be used as an operand of further operations, and it is possible to construct arbitrarily complex queries.
3.1 Constructing regions
The simplest sgrep expression is a text pattern, which evaluates to the set of the regions
that are occurrences of the pattern. The current implementation of sgrep supports text
pattern matching only by searching for either exact or case-insensitive occurrences of
text phrases. For example, the file is searched below for occurrences of the phrases
given between quotes. The -o option is used to specify that the regions are output as
pairs of start and end positions, implementing a sort of inverse operation of the query in
the previous section.
% sgrep -o"(%s,%e)" '"012345" or "23456789" or \
"345678" or "4567" or "456789"' file10
(0,5)(2,9)(3,8)(4,7)(4,9)
We say that a region set is nested, if it contains nested regions. The most important difference between sgrep and other region algebras is that sgrep is capable of manipulating nested region sets. Nested region sets are most often created by the central
“followed-by” operator of sgrep, which is denoted by two dots “..”. The “followedby” operator takes two region sets and creates new regions starting with a region from
the first set and extending to the end of a following region from the second set. The
operation is useful for retrieving complete document elements by their delimiting tags:
% sgrep '"<TITLE>" .. "</TITLE>"' sgrepman.html
<TITLE>sgrep - Manual page</TITLE>
Using sgrep for querying structured text files
5 November 1996
3
Pairs of parentheses are a typical example of nested (or recursive) structures. For an
example, assume that we have the following file containing parenthesized expressions:
% cat example1
( 2+5 ) * ( 10 * ( 7* ( sqr(9) + 3 ) )
Now the “followed-by” operator allows recognizing regions surrounded by balanced
parentheses:
% sgrep -o"-%r-" '"(" .. ")"' example1
-( 2+5 )--( 7* ( sqr(9) + 3 ) )--( sqr(9) + 3 )--(9)The semantics of “..” is defined for arbitrary operand sets as a generalization of the way
how parentheses are matched together “from inside out”. Note that the second opening
parenthesis was excluded from the result above, since there was no matching pair for it.
There are also variations “_.”, “._” and “__” of the “..” operator, which exclude,
correspondingly, either the opening region, or the closing region, or both, from the
resulting regions. For example, the “__” operator can be used to retrieve contents of
document elements excluding their delimiters:
% sgrep '"<TITLE>" __ "</TITLE>"' sgrepman.html
sgrep - Manual page
The operator A extracting B, where A and B are region sets, evaluates to the set of
the regions that result by eliminating from the regions in A any overlap with regions in
B. For example, contents of SGML or HTML documents without any tags could simply be viewed as follows:
% sgrep 'start .. end extracting ("<"..">")' sgrep.html
Sgrep home page
What is sgrep?
From the man page:
...
Sgrep evaluates the queries from left to right, if not indicated otherwise by parentheses.
For example, above it was necessary to indicate that the sub-expression "<"..">"
evaluating to the regions of tags was to be evaluated first.
3.2 Containment conditions
All region algebras provide similar operations for choosing regions either by their contents or by their context. In sgrep these operations are denoted for region sets A and B
by operators A containing B, A not containing B, A in B, and A not in
B.
Using sgrep for querying structured text files
5 November 1996
4
The A containing B operator returns those regions in region set A that contain some
region of region set B. For example, the H2 elements containing a given string could be
retrieved as follows:
% sgrep -o"%r\n" '"<H2>" .. "</H2>" containing "EXPRES"' sgrepman.html
<H2>SYNTAX OF EXPRESSIONS</H2>
<H2>SEMANTICS OF EXPRESSIONS</H2>
The A not containing B operator evaluates to those regions in region set A that do
not contain any region in region set B:
% sgrep -o"%r\n" '"<H2>" .. "</H2>" not containing "EXPRES"' sgrepman.html
<H2>CONTENTS</H2>
<H2>SYNOPSIS</H2>
<H2>DESCRIPTION</H2>
...
The A in B operation evaluates to the regions in region set A that are included in some
region in region set B. The below query retrieves the TITLE-elements of any HTML
file that contains the string “SGML” in the current directory:
% sgrep -o"%r\n" '"<TITLE>".."</TITLE>" in \
(start .. end containing "SGML")' *.html
<TITLE>sgrep - Home page</TITLE>
The A not in B operation returns the regions of A that are not included in any region
of set B. The next command counts, using the -c option, the number of strings
“HREF” not occurring within tags in the HTML files:
%sgrep -c '"HREF" not in ("<" .. ">")' *.html
38
3.3 Removing nested regions
Sometimes there is a need to select the innermost or the outermost regions from a
nested region set, for example to select the document root element or elements consisting of text only. For this purpose there are two functions defined for a region set A:
inner(A) and outer(A). The inner(A) function returns the regions of set A that
are not included in any other region of A. The outer(A) function returns the regions
of set A that do not include any other region of A.
Often a group of regions that cover a single fragment of text is better seen as a single
region. This can be obtained by the concat(A) function, which returns a minimal
Using sgrep for querying structured text files
5 November 1996
5
region set covering the text fragments covered by region set A. That is, for the result
concat chooses the outermost regions of A, and combines any regions that overlap or
start immediately after the preceding one into a single region. By default, if no other
output format is specified, sgrep applies concat to the result before displaying it to
the user.
3.4 Set operations
Sgrep offers typical set operations for forming unions, intersections and differences of
sets. The operations are denoted by the operators A or B, A equal B, and A not
equal B.
The operator A or B evaluates to the set of regions that occur either in region set A or
in region set B, or in both. For example, article elements that discuss either dogs or
cats could be retrieved by the following query:
%sgrep '"<article>" .. "</article>" containing ("dog" or "cat")' doc.sgml
The value of operator A equal B is the set of regions that occur in both region sets A
and B. For example, any second level elements in an SGML document, which contain
the string “HTML” could be obtained as follows:
%sgrep 'outer(ELEMS in ELEMS) equal (ELEMS containing "HTML")’ doc.sgml
The above query refers to a macro called ELEMS, which is a shorthand notation for a
query that evaluates to regions of SGML elements. We will discuss using macros in the
next section.
Finally, the set difference of region sets A and B can be formed using operator A not
equal B.
3.5 Nearness conditions
Sgrep allows combining a given number of subsequent regions of a region set. Operator join(n, A) produces for each region of set A a new region which extends from
the start of that region to the end of the nth region after it, in increasing dictionary order
of the start and end points of regions. The main use of the join operator is in expressing
nearness conditions. The built-in region set chars, whose value is the set of all single
character regions, is useful in this context. For example, the following query would
retrieve articles where “SGML” precedes “database” within 30 characters:
"<article>" .. "</article>" containing ( "SGML" .. "database" in join(30,chars) )
4 Manipulating SGML documents
In this section we show how sgrep can be applied to the manipulation of SGML documents. It should be emphasized that sgrep is not a real SGML tool, since it does not
contain an SGML parser. Sgrep can be made to mimic SGML awareness, but that
requires the documents to be fully tagged, and the queries need to consider the concrete
Using sgrep for querying structured text files
5 November 1996
6
markup syntax and some peculiarities like empty elements. After all, sgrep only picks
text regions satisfying the patterns and conditions in its search query.
Expressing the fetching of SGML elements by sgrep queries directly would often be
too tedious. For this reason sgrep provides a command line option -p for filtering the
queries through an external preprocessor before evaluating them. Currently the m4
macro processor is used for this purpose. Macro definitions can be passed to sgrep in
files using its -f option. Below we show samples of m4 macro definitions that we
have found useful for querying SGML documents. The first eight lines define ready-touse queries for locating markup of tags, and allow using shorthand identifiers like
ELEMS in sgrep queries.
define(ETAGO, "</" )
# End-TAG Open
define(MDO, "<!" )
# Markup Declaration Open
define(PIO, "<?" )
# Proc. Instruction Open
define(STAGO, ("<" not in (ETAGO or MDO or PIO)) ) # Start-TAG Open
define(TAGC, ">" )
# TAG Close
define(STAG, (STAGO .. TAGC) )
# Start TAG
define(ETAG, (ETAGO .. TAGC) )
# End TAG
define(ELEMS, (STAG .. ETAG) )
# all elements
The next macro definitions pertain to recognizing generic identifiers and attribute lists
of elements. These macros allow retrieving values of named attributes by calls like
NAMED_ATTRVAL(status). The BLANK macro evaluates to the regions of single
white space characters.
define(VI, "=")
# value indicator
define(LIT, "\"")
# literal string delimiter
define(ALIT, "'")
# alternative literal string delimiter
define(GI, (STAGO __ (BLANK or TAGC) in STAG)) # Generic Identifier
define(ATTR_LISTS, (GI __ TAGC) )
# All attribute lists
define(ATTR_NAMES, (BLANK __ (VI or BLANK)
in (BLANK __ (VI or BLANK) .. VI in ATTR_LISTS)) )
define(ATTRS, (ATTR_NAMES ._ (ATTR_NAMES or TAGC)) )
define(NAMED_ATTRS,(ATTRS containing ("$1" equal ATTR_NAMES)) )
define(NAMED_ATTRVAL, ((VI __ (ATTR_NAMES or TAGC)
Using sgrep for querying structured text files
5 November 1996
7
not containing (LIT or ALIT) extracting BLANK)
or (LIT __ LIT) or (ALIT __ ALIT)
in (NAMED_ATTRS($1) extracting ATTR_NAMES)) )
The next macro definitions allow sgrep queries to retrieve element tags, elements and
their contents by the GI (generic identifier) of the element:
define(NAMED_STAG,( "<$1>" or ( ("<$1 " or "<$1\t" or "<$1\n") .. ">")) )
define(NAMED_ELEMS,(NAMED_STAG($1) .. "</$1>") )
define(NAMED_CONTENTS,( NAMED_STAG($1) __ "</$1>" ))
Writing such macro definitions is subtle, but once they have been written and tested,
they make expressing high-level queries quite easy. For example, if an SGML document file contains book elements that express the number of pages in the book by a
pages attribute and the price of the book in a price element, the following query
would find the books that have 200 pages and cost $20:
NAMED_ELEMS(book) containing
(NAMED_ATTRVAL(pages) equal "200" in NAMED_STAG(book))
containing (NAMED_CONTENTS(price) equal "$20")
The below command counts the total number of chapters in “The SGML Handbook”,
when the title of a book is given in its name element:
sgrep -c 'NAMED_ELEMS(chapter) in
\
(NAMED_ELEMS(book) containing \
(NAMED_CONTENTS(name) equal "The SGML Handbook"))'
Similarly books could be retrieved by the contents of their author element:
NAMED_ELEMS(book) containing
(NAMED_ELEMS(author) containing "Herwijnen")
4.1 Applying sgrep to document assembly
An on-going research and development project called SID (Structured and Intelligent
Documents) is concerned with adding intelligent features to documents in order to
make their reuse and assembly more efficient [AHH96]. The SID project participates in
the development of a dynamic collection of text books for the National Board of Education. A goal is to allow easy production of tailored high-quality versions of the books
in the collection. To enable this the books have been converted from a Word format to a
Using sgrep for querying structured text files
5 November 1996
8
slightly modified ISO 12083 book DTD. The modifications include the enrichment of
books with attributes for specifying, e.g., the didactic form (introductory, base material,
advanced, applied, etc.) of the contents of document elements.
We have developed an architecture which allows tailored versions of the books to be
created by filling an order form, specifying in a table of contents which parts of a book
are to be chosen to the assembly, and possibly restricting the chosen contents to be of
appropriate didactic form. (See Figure 4.1.)
Figure 4.1 The assembly architecture
The crux of the architecture is the compilation of the order form into a query that
selects and combines the document fragments form the collection. As the current solution we translate the form into an sgrep query. The translation is based on the following
principles. First, the query generator creates a definition of a macro called
BASE_FRAGMENTS, which evaluates to the fragments chosen by the form. A chapter
selected and restricted to a didactic form in the order form is translated into a subexpression of BASE_FRAGMENTS, roughly of the following form:
or CHAP_BY_ID(<Id of the chosen chapter>)
extracting P_NOT_OF_CONTENT(<didactic form chosen>)
These expressions refer to sgrep macros written specially for the material in the collection. The macros, written in upper-case above, allow retrieving chapters and sections
by their ID attribute values, and paragraphs by the attribute value for their didactic
form. The definitions of these macros are in principle similar to the macro definitions
we have seen above.
Similarly, sub-queries for the regions of each selected section are added to the query
macro. Then some additional parts are added to the result query, both to wrap the book
fragments into complete SGML elements and to add introductory context like initial
text paragraphs to the sections retrieved in the assembly:
NAMED_STAG(chapter) in (CHAPTERS containing BASE_FRAGMENTS)
. _ NAMED_STAG(section)
Using sgrep for querying structured text files
5 November 1996
9
or NAMED_ETAG(chapter) in (CHAPTERS containing BASE_FRAGMENTS)
Finally, the front and back materials of the original books are included in the assembly.
The generated sgrep query creates a digest of the original SGML documents by retrieving their fragments in order. When the fragments are picked carefully, by applying the
principles outlined above, the result is a valid SGML instance, which can be loaded
either to a browser or to an editor. We have tested the prototype of the architecture
using the Softquad Panorama browser. The production of book-quality layouts is
planned to be done by loading the results in FrameMaker+SGML.
5 Conclusion
We have described the main features of the sgrep query language, which implements an
unlimited algebra of text regions. Even though the algebra and its implementation give
no interpretation to the contents of documents, we have seen above that sgrep can be
used for rather high-level manipulation of, say, SGML documents.
Sgrep has been developed and tested in Linux and Unix systems. The software, currently in version 0.99, is freely available at ftp://ftp.cs.helsinki.fi/pub/
Software/Local/Sgrep/.
Biography: Jani Jaakkola. Jani Jaakkola is a full-time student of computer science at
the University of Helsinki. Currently he works as a systems designer in the SID (Structured and Intelligent Documents) project.
Biography: Pekka Kilpeläinen. Pekka Kilpeläinen is an assistant professor of computer science at the University of Helsinki. He received his master's degree in 1989, the
degree of the licentiate of philosophy in 1992, and his Ph.D. in 1993, all in computer
science at the University of Helsinki. Currently he acts as the project manager in the
SID project.
References
[AHH96] Constructing tailored SGML documents. H. Ahonen, B. Heikkinen, O. Heinonen, J. Jaakkola, P. Kilpeläinen, G. Lindén and H. Mannila. This volume. 1996.
[AKW88] The AWK Programming Language. A. V. Aho, B. W. Kernighan and P. J.
Weinberger. Addison-Wesley. 1988.
[Bur92] An Algebra for Hierarchically Organized Text-Dominated Databases. F. J.
Burkowski. Information Processing & Management. 1992. (333-348)
[CCB95] An Algebra for Structured Text Search and a Framework for its Implementation. C.L.A. Clarke, G.V. Cormack and F. J. Burkowski. The Computer Journal. 1995.
(43-56)
[JK96] Sgrep - A Tool to Search Structured Text. J. Jaakkola and P. Kilpeläinen. University of Helsinki. (In preparation) 1996.
Using sgrep for querying structured text files
5 November 1996
10
[SaT92] PAT Expressions: an Algebra for Text Search. A. Salminen and F. Wm.
Tompa. UW Centre for the New Oxford English Dictionary and Text Research report
OED-92-02. 1992.
[ThR75] UNIX Programmer's Manual. K. Thompson and D.M. Ritchie. Bell Laboratories. 1975.
[WaS91] Programming Perl. L. Wall and R.L. Schwartz. O’Reilly & Associates, Inc..
1991.
Using sgrep for querying structured text files
5 November 1996
11
Download